Beyond Uniform Lipschitz Condition in Differentially Private Optimization

Authors: Rudrajit Das, Satyen Kale, Zheng Xu, Tong Zhang, Sujay Sanghavi

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify the efficacy of our recommendation via experiments on 8 datasets. In Section 5.2, we corroborate our recommendation with experiments on four1 vision datasets, viz., Caltech-256 (Griffin et al., 2007), Food-101 (Bossard et al., 2014), CIFAR-100 and CIFAR-10, and two language datasets, viz., Tweet Eval Emoji (Barbieri et al., 2018) and Emotion (Saravia et al., 2018).
Researcher Affiliation Collaboration Rudrajit Das * 1 Satyen Kale 2 Zheng Xu 2 Tong Zhang 2 3 Sujay Sanghavi 1 1UT Austin 2Google Research 3HKUST.
Pseudocode Yes Algorithm 1 DP-SGD (Abadi et al., 2016)
Open Source Code No The paper mentions using a third-party library ('Py Torch s Opacus library') but does not state that the authors' own implementation code for the described methodology is open-source or provide a link to it.
Open Datasets Yes Our experiments here are conducted on four vision datasets available in Torchvision, viz., Caltech-256 (Griffin et al., 2007), Food-101 (Bossard et al., 2014), CIFAR-100 and CIFAR-10, and two language datasets, viz., Tweet Eval Emoji (Barbieri et al., 2018) and Emotion (Saravia et al., 2018).
Dataset Splits No The paper refers to 'test accuracy' and hyperparameter tuning but does not specify the exact train/validation/test split percentages, sample counts, or methodology used for data partitioning.
Hardware Specification Yes A single NVIDIA TITAN Xp GPU was used for all the experiments in this paper.
Software Dependencies No Py Torch s Opacus library (Yousefpour et al., 2021) is used for private training. (No version numbers provided for PyTorch or Opacus).
Experiment Setup Yes We consider three privacy levels (2, 10 5)-DP, (4, 10 5)-DP and (6, 10 5)-DP, with batch size = 500. We test several values of the clip norm τ, viz., the 0th, 10th, 20th, 40th, 80th and 100th percentile of the per-sample Lipschitz constants... For each value of τ, we tune over several values of the constant learning rate η, viz., {0.0001, 0.0003, 0.0006, 0.001, 0.003, 0.006, 0.01, 0.03, 0.06, 0.1, 0.3, 0.6, 1, 3, 6, 10}.